← Back to directory
C

CodeInspectus MCP Server

Community
Local-first MCP security scanner for AI-generated apps
Category
Dev Tools #5 of 438
Stars
★ 48 Popular
Transport
stdio (local process)
Runtime
Node.js 22+ · Prebuilt binary cosign on PATH for O
Credentials
No credential needed
License
Apache-2.0
Last commit
Tools
6
75FMRS · B

An unusually disciplined local security scanner: SHA-pinned engines, zero scan-time egress, no telemetry, redacted secrets, a three-state honest evidence model, and rare candor about its own limits (narrow native rules, AI-drafted compliance mappings, heuristic prompt-injection detection). Well suited to agent-driven scan-fix-rescan security loops; not a substitute for deep audits, runtime testing, or compliance certification. MIT licensed, solo-maintained by Synvoya, with independent security review actively solicited.

Strongest · Documentation 19/20 Weakest · Reliability 11/20

Reliability
11/20
Security and permissions
18/20
Maintenance
13/20
Documentation
19/20
Setup experience
14/20
Why each score
Reliability 11/20
The manifest's single npm stdio package and declared tool list match the README's tool table one-for-one, and version 3.2.0 plus a long documented changelog history (v0.3.1→v3.2.0) suggest the described happy path is plausible. However, this is a text-only review — no actual CI workflow YAML, test files, or eval harness output were supplied for verification, only README references to 'npm run eval' and E23-E52 eval IDs. Per static calibration, reliability cannot exceed 12 without directly-inspected test/CI evidence.
Security and permissions 18/20
The design shows strong least-privilege discipline: read-only scanning by default ('never edits or deletes your source code'), redacted secret previews, SHA-pinned and cosign-verified engine binaries with fail-closed verification, explicit approval gates before any fix or git operation, and disclosed zero-egress-at-scan-time with clearly scoped exceptions (explicit repair/setup commands only). No red-line behavior (no silent destructive defaults, no credential exfiltration, no fabricated claims) appears in the supplied material. Points withheld because all of this is self-reported in an untrusted README/manifest with no actual source code available to verify the claims independently.
Maintenance 13/20
Version 3.2.0 with an extensive documented release history (v0.3.1, 0.3.2, v1.5.0, v2.1.0, v3.2.0) and a maintained CHANGELOG.md indicate active, iterative development; MIT license is present; open issues (5) and 'good first issue' links show some community engagement. However, this is explicitly a solo-maintainer project ('no company behind it'), with only 37 stars, and no dedicated SECURITY.md or documented response-time/disclosure channel was shown, leaving governance and bus-factor risk.
Documentation 19/20
Documentation is exceptionally layered: install prerequisites, per-client registration snippets (Claude Code, Cursor, VS Code, Codex, Windsurf/Cline/Aider), a full CLI command reference, an explicit tool table with I/O behavior, and an unusually thorough 'Honest claims' section enumerating exact scope/limits per rule and pack (e.g., which languages, which frameworks, which non-detections). Numerous doc links (CI-POLICY, BASELINES-AND-TRIAGE, SCAN-HISTORY, COMPLIANCE-RATIONALE, etc.) are referenced for deeper detail. Slight deduction because this depth cannot be cross-checked against the linked doc files themselves in this static, README-only review.
Setup experience 14/20
Setup is documented with a clear few-step flow (Node ≥22, optional cosign, `npx codeinspectus repair-engines`) and exact JSON/TOML config snippets for six different clients, plus an explicit `verify-engines` re-check and clear `engine_setup` state reporting for troubleshooting. Per static calibration, setup is capped at 15 without inspectable CI/test evidence proving the install path actually works end-to-end; the score reflects strong documentation of the process rather than confirmed execution.

Static review · not runListed 2026-09-07

Read the FMRS scoring method →

Fit and risk

What it can accessReads local filesWrites / deletes local filesRuns commands or code

Best for

  • Security-conscious developers building apps with AI agents like Claude Code, Cursor, or Codex
  • Individuals and small teams who want fully offline, local code scanning with no source upload
  • Teams mapping CWE-keyed findings to code-visible compliance controls

Not for

  • Teams seeking full compliance certification or runtime security proof (the tool explicitly does neither)
  • Deep audits requiring cross-file/whole-program dataflow or runtime mobile testing
  • Restricted environments that cannot run Node.js ≥22 or install cosign and engine binaries
  • Essential Eight assessments (only ~1 of 8 mitigations is code-evidenced)

Required permissions

  • Reads project source and repository files (read-only; never edits or deletes your code)
  • Stores engine data and scan history under ~/.codeinspectus
  • Optionally writes one SBOM file (managed directory by default, or a chosen path)
  • Runs SHA-verified Opengrep/Gitleaks/Trivy binaries as local subprocesses
  • Reads git state for a git-safety recommendation; git operations require user approval and the tool never runs them itself

Risks and side effects

  • Prompt-injection detection is heuristic and immature; findings are worded "potential …" at medium confidence
  • Fixes are applied by your AI agent and need human review; the tool only reports and never edits code
  • Native rule coverage is deliberately narrow (mostly intrafile analysis); do not infer broader coverage than executed packs report
  • Compliance mappings are AI-drafted and maintainer-reviewed with 0 of 96 community-verified — not an audit
  • Engine binaries carry supply-chain risk, mitigated via SHA pinning, though setup/repair still downloads over the network

Setup

Before you start

Runtime:Node.js 22+ · Prebuilt binary cosign on PATH for O

Prerequisites: Node.js ≥22 (Node 24 LTS recommended); cosign on PATH when Opengrep or Trivy binaries need installation (fail-closed verification; Gitleaks verifies by checksum and needs no cosign). Register once per machine with your client (e.g. Claude Code: claude mcp add- codeinspectus '{"command":"npx","args":["-y","codeinspectus"]}'; Cursor: add to mcp.; VS Code: code --add-mcp; Codex: codex mcp add or config.toml with tool_timeout_sec = 600 recommended). Then run npx codeinspectus repair-engines to install hash-verified engines and databases. Scans then perform zero network I/O. Re-verify binaries any time with npx codeinspectus verify-engines.

.mcp.json
{"mcpServers":{"codeinspectus":{"command":"npx","args":["-y","codeinspectus"]}}}

Shown for Claude Code. Other clients may use a different file or key (VS Code uses "servers") — the configurator below converts it.

.vscode/mcp.json
{
  "servers": {
    "codeinspectus": {
      "command": "npx",
      "args": [
        "-y",
        "codeinspectus"
      ]
    }
  }
}

Goes in your project's .vscode/mcp.json (VS Code uses a "servers" key).

Terminal
claude mcp add codeinspectus -- npx -y codeinspectus

Run it in a terminal; replace any <…> placeholders with your own values first.

Check that it works

After registering the client, run npx codeinspectus repair-engines once to install engines, then confirm the six tools (codeinspectus_scan, codeinspectus_rescan, codeinspectus_compliance_report, codeinspectus_explain_finding, codeinspectus_generate_sbom, codeinspectus_list_rules) appear in the client's tool list; calling codeinspectus_list_rules should show active detectors and engine_setup state.

Troubleshooting

  1. Scan returns engine_setup other than ready: have the agent explain the state (repair_required / db_refresh_recommended / unsupported_platform) and, with approval, run npx codeinspectus repair-engines
  2. CVE rescan tracking not enabled: for a Trivy DB installed before 0.3.2, run repair-engines once to re-fetch via the verified path
  3. Client timeouts (especially Codex): set tool_timeout_sec to 600, since large-repository scans can exceed the default 60 seconds
  4. Native pack did not run: check pack_coverage and detected_technologies — each pack requires exact framework/dependency evidence (e.g. Flutter, openai-php/client, async-openai)
  5. Verify binary integrity: npx codeinspectus verify-engines; unpinned or mismatched binaries are refused

Things to try

Once connected, you can ask your AI assistant things like:

  • Scan the current project for security vulnerabilities and list all CWE-keyed findings
  • I fixed some issues — rescan and diff against the previous scan to show resolved, remaining, and introduced findings
  • Generate a CycloneDX SBOM for this project
  • Explain finding CI-0001 in detail with its full remediation steps

Tools 6

codeinspectus_scan read-only
Full local scan of a path (engines + AI checks). Returns CWE-keyed findings, detected technologies, native-pack and Pub dependency coverage, remediations, framework tags, and three-state repository evidence for supported runtime controls.
codeinspectus_rescan read-only
Re-scan after fixes; diffs against a prior scan producing resolved / remaining / introduced, with fresh technology and pack coverage.
codeinspectus_compliance_report read-only
Per-framework code-level control coverage report (not certification).
codeinspectus_explain_finding read-only
Deep explanation plus full remediation for a single finding.
codeinspectus_generate_sbom writes
Generate a CycloneDX/SPDX SBOM using Trivy plus native Pub inventory/fallback (written to a managed directory by default, or a path you choose).
codeinspectus_list_rules read-only
List active detectors, native-pack inventory and rule ownership, engine versions, detection-DB provenance and freshness, and structured setup/repair state.

Use cases

Scanning AI-generated / vibe-coded apps for real vulnerabilities before shipping
Agent-driven scan → fix → rescan security loops
Producing code-level control coverage reports for OWASP, NIST CSF, ISO 27001, SOC 2 and more
Generating local SBOMs (CycloneDX/SPDX) including Dart/Flutter pubspec.lock dependency inventory
CI/local baseline and history comparisons, sealed evidence bundles, and GitHub/Jira/Linear issue payload export

Supported clients

Claude Code
Cursor
VS Code
Codex
Windsurf
Cline
Aider

Listed from the project's documentation, not tested by this site.

Overview

CodeInspectus is a local-first, privacy-preserving MCP security scanner and CLI maintained by Synvoya. Any AI coding agent (Claude Code, Cursor, Codex, Windsurf, Cline, Aider) can invoke it to scan AI-generated / "vibe-coded" code for real vulnerabilities, map findings to code-level compliance-framework coverage, and drive a scan → fix → rescan loop — fully on your machine, no account, zero network egress at scan time. It orchestrates three OSS engines — Opengrep (SAST/OWASP Top 10), Gitleaks (secrets), and Trivy (dependency CVEs, IaC, SBOM) — plus first-party AI-code-specific checks (Supabase RLS / inverted authz, prompt-injection sinks, model tool arguments reaching shell sinks, client-writable user_metadata authorization, dangerous HTML rendering, Firebase public writes, GitHub Actions workflow risks, and more), shipping 88 curated detections normalized into a CWE-keyed schema. Engine binaries are SHA-pinned and signature-verified; there is no telemetry. Compliance mappings are code-level coverage only, never certification.

Similar servers

Context7 80 · B

Upstash's official server providing up-to-date third-party library docs for AI coding assistants

★ 63.3k · Tools 2 Compare with this →

Source revision 8dbdd6bc3bfa Data synced 2026-10-19 Read the FMRS scoring method